File size: 34,445 Bytes
4c46869 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 | """
PrivaMesh Legal โ Fine-tuning script
Fine-tune Mistral-7B for legal PII token classification (BIOES scheme)
Usage:
python train.py --config configs/legal_fr.yaml
Requirements:
pip install transformers datasets peft accelerate bitsandbytes seqeval torch
"""
import os
import json
import argparse
import logging
from dataclasses import dataclass, field
from typing import Optional, List, Dict, Any
import torch
import numpy as np
from datasets import Dataset, DatasetDict, load_dataset
from transformers import (
AutoTokenizer,
AutoModelForTokenClassification,
TrainingArguments,
Trainer,
DataCollatorForTokenClassification,
EarlyStoppingCallback,
)
from peft import (
LoraConfig,
TaskType,
get_peft_model,
prepare_model_for_kbit_training,
)
from transformers import BitsAndBytesConfig
import evaluate
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# 1. Label definitions โ BIOES scheme over 24 legal privacy categories
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
PRIVACY_CATEGORIES = [
# Natural persons
"PERSON_NAME",
"LEGAL_COUNSEL",
"JUDGE_NAME",
"SIGNATORY",
"WITNESS",
# Legal entities
"COMPANY_NAME",
"COMPANY_ID",
"COURT_NAME",
"BAR_ASSOCIATION",
# Financial & contractual
"CONTRACT_AMOUNT",
"BANK_ACCOUNT",
"PENALTY_AMOUNT",
# Contact & location
"PRIVATE_ADDRESS",
"PRIVATE_EMAIL",
"PRIVATE_PHONE",
# Temporal & reference
"CONTRACT_DATE",
"DEADLINE",
"CASE_NUMBER",
# RGPD/Compliance specific
"DATA_SUBJECT",
"DPO_IDENTITY",
"PROCESSING_PURPOSE",
"AUDIT_REFERENCE",
"REGULATORY_BODY",
"DIRIGEANT",
]
# Build BIOES label list
LABELS = ["O"]
for cat in PRIVACY_CATEGORIES:
for prefix in ["B", "I", "E", "S"]:
LABELS.append(f"{prefix}-{cat}")
LABEL2ID = {label: idx for idx, label in enumerate(LABELS)}
ID2LABEL = {idx: label for label, idx in LABEL2ID.items()}
NUM_LABELS = len(LABELS) # 97
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# 2. Placeholder mapping โ semantic replacements per category
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
PLACEHOLDER_TEMPLATES = {
"PERSON_NAME": "[PERSONNE_{n}]",
"LEGAL_COUNSEL": "[AVOCAT_{n}]",
"JUDGE_NAME": "[MAGISTRAT_{n}]",
"SIGNATORY": "[SIGNATAIRE_{n}]",
"WITNESS": "[TEMOIN_{n}]",
"COMPANY_NAME": "[SOCIETE_{n}]",
"COMPANY_ID": "[SIRET_{n}]",
"COURT_NAME": "[JURIDICTION_{n}]",
"BAR_ASSOCIATION": "[BARREAU_{n}]",
"CONTRACT_AMOUNT": "[MONTANT_{n}]",
"BANK_ACCOUNT": "[IBAN_{n}]",
"PENALTY_AMOUNT": "[PENALITE_{n}]",
"PRIVATE_ADDRESS": "[ADRESSE_{n}]",
"PRIVATE_EMAIL": "[EMAIL_{n}]",
"PRIVATE_PHONE": "[TEL_{n}]",
"CONTRACT_DATE": "[DATE_{n}]",
"DEADLINE": "[ECHEANCE_{n}]",
"CASE_NUMBER": "[DOSSIER_{n}]",
"DATA_SUBJECT": "[PERSONNE_CONCERNEE_{n}]",
"DPO_IDENTITY": "[DPO_{n}]",
"PROCESSING_PURPOSE": "[FINALITE_{n}]",
"AUDIT_REFERENCE": "[AUDIT_REF_{n}]",
"REGULATORY_BODY": "[AUTORITE_{n}]",
"DIRIGEANT": "[DIRIGEANT_{n}]",
}
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# 3. Synthetic data generator โ creates training examples
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
class LegalDataGenerator:
"""
Generates synthetic annotated legal document examples for training.
Used to bootstrap training data before fine-tuning on real corpora.
"""
PERSON_NAMES_FR = [
"Jean Dupont", "Marie Martin", "Pierre Leblanc", "Sophie Durand",
"Ahmed Benali", "Fatima Zahra", "Karim Mansouri", "Isabelle Lefebvre",
"Franรงois Moreau", "Nathalie Petit", "Mehdi Rachidi", "Claire Rousseau",
]
COMPANY_NAMES = [
"Nexum SAS", "TechLegal SA", "DataPro SARL", "InnovateFR SAS",
"ConsultPro SAS", "LegalTech SA", "SecureData SARL", "CloudFR SAS",
]
LAWYERS = [
"Maรฎtre Jean Dupont", "Maรฎtre Sophie Martin", "Maรฎtre Ahmed Benali",
"Maรฎtre Claire Rousseau", "Maรฎtre Franรงois Moreau",
]
AMOUNTS = [
"150 000 EUR", "25 000 EUR", "500 000 EUR", "75 000 EUR",
"12 500 EUR", "1 000 000 EUR", "350 000 EUR",
]
SIRETS = [
"123 456 789 00012", "987 654 321 00045", "456 789 123 00078",
"321 654 987 00034",
]
ADDRESSES = [
"12 rue de la Paix, 75001 Paris",
"45 avenue Victor Hugo, 69002 Lyon",
"8 place de la Rรฉpublique, 33000 Bordeaux",
"22 boulevard Haussman, 75009 Paris",
]
EMAILS = [
"j.dupont@cabinet-dupont.fr",
"contact@nexumsas.fr",
"direction@techpro.com",
]
IBANS = [
"FR76 3000 4000 0100 0000 1234 567",
"FR76 1670 6000 0302 0060 0800 073",
]
TEMPLATES = [
{
"text": "Le contrat conclu entre {lawyer}, avocat au barreau de Paris (SIRET {siret}), et la sociรฉtรฉ {company}, reprรฉsentรฉe par {person} en qualitรฉ de Directeur Gรฉnรฉral, prรฉvoit une indemnitรฉ de rupture de {amount} conformรฉment ร l'article L.1237-19 du Code du travail.",
"entities": [
("lawyer", "LEGAL_COUNSEL"),
("siret", "COMPANY_ID"),
("company", "COMPANY_NAME"),
("person", "DIRIGEANT"),
("amount", "CONTRACT_AMOUNT"),
]
},
{
"text": "M. {person}, domiciliรฉ au {address}, a mandatรฉ {lawyer} pour le reprรฉsenter dans la procรฉdure RG nยฐ24/01234 devant le Tribunal de Commerce de Paris.",
"entities": [
("person", "PERSON_NAME"),
("address", "PRIVATE_ADDRESS"),
("lawyer", "LEGAL_COUNSEL"),
]
},
{
"text": "Conformรฉment au RGPD, la sociรฉtรฉ {company} (SIRET {siret}) dรฉsigne {person} en qualitรฉ de Dรฉlรฉguรฉ ร la Protection des Donnรฉes (DPO), joignable ร l'adresse {email}.",
"entities": [
("company", "COMPANY_NAME"),
("siret", "COMPANY_ID"),
("person", "DPO_IDENTITY"),
("email", "PRIVATE_EMAIL"),
]
},
{
"text": "La prรฉsente convention de prestation de services est conclue entre {company1} (RCS Paris B {siret}) et {company2}, reprรฉsentรฉe par {person}, pour un montant annuel de {amount} HT, payable par virement sur le compte {iban}.",
"entities": [
("company1", "COMPANY_NAME"),
("siret", "COMPANY_ID"),
("company2", "COMPANY_NAME"),
("person", "DIRIGEANT"),
("amount", "CONTRACT_AMOUNT"),
("iban", "BANK_ACCOUNT"),
]
},
{
"text": "Dans le cadre de l'audit ISO 27001 rรฉfรฉrencรฉ AUD-2024-042, {person} (DPO de {company}) a transmis ร {lawyer} l'ensemble des registres de traitement pour vรฉrification de conformitรฉ avant le {date}.",
"entities": [
("person", "DPO_IDENTITY"),
("company", "COMPANY_NAME"),
("lawyer", "LEGAL_COUNSEL"),
("date", "DEADLINE"),
]
},
]
import random as _random
def generate(self, n_samples: int = 1000) -> List[Dict[str, Any]]:
"""Generate n_samples synthetic annotated examples."""
import random
samples = []
for _ in range(n_samples):
template = random.choice(self.TEMPLATES)
text_template = template["text"]
entity_map = template["entities"]
# Sample values
values = {
"lawyer": random.choice(self.LAWYERS),
"person": random.choice(self.PERSON_NAMES_FR),
"company": random.choice(self.COMPANY_NAMES),
"company1": random.choice(self.COMPANY_NAMES),
"company2": random.choice(self.COMPANY_NAMES),
"siret": random.choice(self.SIRETS),
"amount": random.choice(self.AMOUNTS),
"address": random.choice(self.ADDRESSES),
"email": random.choice(self.EMAILS),
"iban": random.choice(self.IBANS),
"date": "30 juin 2025",
}
# Fill template
text = text_template.format(**values)
# Build character-level annotations
annotations = []
for field_name, label in entity_map:
if field_name not in values:
continue
value = values[field_name]
start = text.find(value)
if start == -1:
continue
annotations.append({
"start": start,
"end": start + len(value),
"label": label,
"text": value
})
samples.append({
"text": text,
"annotations": annotations
})
return samples
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# 4. Tokenization & alignment
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def tokenize_and_align_labels(
examples: Dict[str, Any],
tokenizer,
max_length: int = 512,
label_all_tokens: bool = False,
) -> Dict[str, Any]:
"""
Tokenize text and align BIOES labels with subword tokens.
Handles the subword alignment problem for NER with BIOES scheme.
"""
tokenized = tokenizer(
examples["tokens"],
truncation=True,
max_length=max_length,
is_split_into_words=True,
padding="max_length",
)
aligned_labels = []
for i, label_ids in enumerate(examples["labels"]):
word_ids = tokenized.word_ids(batch_index=i)
previous_word_idx = None
label_ids_aligned = []
for word_idx in word_ids:
if word_idx is None:
label_ids_aligned.append(-100) # special tokens
elif word_idx != previous_word_idx:
label_ids_aligned.append(label_ids[word_idx])
else:
# Subword token of the same word
if label_all_tokens:
label_ids_aligned.append(label_ids[word_idx])
else:
label_ids_aligned.append(-100)
previous_word_idx = word_idx
aligned_labels.append(label_ids_aligned)
tokenized["labels"] = aligned_labels
return tokenized
def char_annotations_to_token_labels(
text: str,
annotations: List[Dict],
tokenizer,
) -> Dict[str, Any]:
"""
Convert character-level annotations to word-level BIOES token labels.
"""
words = text.split()
word_labels = ["O"] * len(words)
# Map character positions to word indices
char_to_word = {}
char_pos = 0
for word_idx, word in enumerate(words):
for _ in word:
char_to_word[char_pos] = word_idx
char_pos += 1
char_pos += 1 # space
# Apply BIOES labels
for ann in annotations:
start_word = char_to_word.get(ann["start"])
end_char = ann["end"] - 1
end_word = char_to_word.get(end_char, start_word)
if start_word is None:
continue
label = ann["label"]
span_words = list(range(start_word, end_word + 1))
if len(span_words) == 1:
word_labels[span_words[0]] = f"S-{label}"
else:
word_labels[span_words[0]] = f"B-{label}"
for w in span_words[1:-1]:
word_labels[w] = f"I-{label}"
word_labels[span_words[-1]] = f"E-{label}"
return {
"tokens": words,
"labels": [LABEL2ID.get(l, 0) for l in word_labels],
}
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# 5. Metrics
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
seqeval = evaluate.load("seqeval")
def compute_metrics(eval_pred):
logits, labels = eval_pred
predictions = np.argmax(logits, axis=-1)
true_predictions = [
[ID2LABEL[p] for p, l in zip(prediction, label) if l != -100]
for prediction, label in zip(predictions, labels)
]
true_labels = [
[ID2LABEL[l] for p, l in zip(prediction, label) if l != -100]
for prediction, label in zip(predictions, labels)
]
results = seqeval.compute(
predictions=true_predictions,
references=true_labels,
scheme="BIOES",
mode="strict",
)
return {
"precision": results["overall_precision"],
"recall": results["overall_recall"],
"f1": results["overall_f1"],
"accuracy": results["overall_accuracy"],
}
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# 6. Model loader โ Mistral + LoRA
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def load_model_and_tokenizer(
base_model: str = "mistralai/Mistral-7B-v0.1",
use_4bit: bool = True,
lora_r: int = 16,
lora_alpha: int = 32,
lora_dropout: float = 0.05,
):
"""Load Mistral with 4-bit quantization + LoRA for efficient fine-tuning."""
tokenizer = AutoTokenizer.from_pretrained(
base_model,
padding_side="right",
add_eos_token=True,
)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
# 4-bit quantization config
bnb_config = BitsAndBytesConfig(
load_in_4bit=use_4bit,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_use_double_quant=True,
) if use_4bit else None
# Load base model for token classification
model = AutoModelForTokenClassification.from_pretrained(
base_model,
num_labels=NUM_LABELS,
id2label=ID2LABEL,
label2id=LABEL2ID,
quantization_config=bnb_config,
device_map="auto",
torch_dtype=torch.bfloat16,
)
if use_4bit:
model = prepare_model_for_kbit_training(model)
# LoRA configuration
lora_config = LoraConfig(
task_type=TaskType.TOKEN_CLS,
r=lora_r,
lora_alpha=lora_alpha,
lora_dropout=lora_dropout,
bias="none",
target_modules=[
"q_proj", "v_proj", "k_proj", "o_proj",
"gate_proj", "up_proj", "down_proj",
],
modules_to_save=["classifier"], # always train the classification head
)
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
return model, tokenizer
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# 7. Dataset preparation
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def prepare_dataset(
tokenizer,
n_synthetic: int = 5000,
hf_dataset: Optional[str] = None,
max_length: int = 512,
) -> DatasetDict:
"""
Prepare training dataset.
Combines synthetic data + optional HuggingFace dataset.
"""
generator = LegalDataGenerator()
raw_samples = generator.generate(n_samples=n_synthetic)
# Convert to token-level format
token_samples = []
for sample in raw_samples:
converted = char_annotations_to_token_labels(
sample["text"],
sample["annotations"],
tokenizer,
)
token_samples.append(converted)
# Split train/val/test
n = len(token_samples)
n_train = int(n * 0.80)
n_val = int(n * 0.10)
dataset = DatasetDict({
"train": Dataset.from_list(token_samples[:n_train]),
"validation": Dataset.from_list(token_samples[n_train:n_train + n_val]),
"test": Dataset.from_list(token_samples[n_train + n_val:]),
})
# Tokenize and align
tokenized = dataset.map(
lambda x: tokenize_and_align_labels(x, tokenizer, max_length),
batched=True,
remove_columns=["tokens", "labels"],
)
logger.info(f"Dataset sizes โ train: {len(tokenized['train'])}, "
f"val: {len(tokenized['validation'])}, "
f"test: {len(tokenized['test'])}")
return tokenized
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# 8. Training
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
def train(
base_model: str = "mistralai/Mistral-7B-v0.1",
output_dir: str = "./privamesh-legal-output",
n_synthetic: int = 5000,
num_epochs: int = 5,
batch_size: int = 4,
gradient_accumulation: int = 4,
learning_rate: float = 2e-4,
max_length: int = 512,
use_4bit: bool = True,
push_to_hub: bool = False,
hub_model_id: str = "sallani/PrivaMesh",
):
logger.info("=" * 60)
logger.info("PrivaMesh Legal โ Fine-tuning Mistral-7B")
logger.info(f"Base model : {base_model}")
logger.info(f"Output dir : {output_dir}")
logger.info(f"Labels : {NUM_LABELS} ({len(PRIVACY_CATEGORIES)} categories ร BIOES + O)")
logger.info("=" * 60)
# Load model
model, tokenizer = load_model_and_tokenizer(
base_model=base_model,
use_4bit=use_4bit,
)
# Prepare data
dataset = prepare_dataset(
tokenizer=tokenizer,
n_synthetic=n_synthetic,
max_length=max_length,
)
data_collator = DataCollatorForTokenClassification(
tokenizer=tokenizer,
padding=True,
max_length=max_length,
)
# Training arguments
training_args = TrainingArguments(
output_dir=output_dir,
num_train_epochs=num_epochs,
per_device_train_batch_size=batch_size,
per_device_eval_batch_size=batch_size,
gradient_accumulation_steps=gradient_accumulation,
learning_rate=learning_rate,
lr_scheduler_type="cosine",
warmup_ratio=0.05,
weight_decay=0.01,
evaluation_strategy="epoch",
save_strategy="epoch",
load_best_model_at_end=True,
metric_for_best_model="f1",
greater_is_better=True,
logging_dir=f"{output_dir}/logs",
logging_steps=50,
report_to="none",
bf16=True,
tf32=True,
dataloader_num_workers=4,
group_by_length=True,
push_to_hub=push_to_hub,
hub_model_id=hub_model_id if push_to_hub else None,
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset["train"],
eval_dataset=dataset["validation"],
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
callbacks=[EarlyStoppingCallback(early_stopping_patience=2)],
)
# Train
logger.info("Starting training...")
trainer.train()
# Evaluate on test set
logger.info("Evaluating on test set...")
test_results = trainer.evaluate(dataset["test"])
logger.info(f"Test results: {json.dumps(test_results, indent=2)}")
# Save
logger.info(f"Saving model to {output_dir}")
trainer.save_model(output_dir)
tokenizer.save_pretrained(output_dir)
# Save config with PrivaMesh metadata
config_path = os.path.join(output_dir, "config.json")
with open(config_path) as f:
config = json.load(f)
config["privamesh"] = {
"version": "1.0.0",
"domain": "legal",
"languages": ["fr", "en"],
"regulatory_coverage": ["RGPD", "DORA", "NIS2", "ISO27001", "ISO42001"],
"mesh_role": "specialist",
"test_f1": test_results.get("eval_f1", 0.0),
}
with open(config_path, "w") as f:
json.dump(config, f, indent=2)
if push_to_hub:
logger.info(f"Pushing to HuggingFace Hub: {hub_model_id}")
trainer.push_to_hub()
logger.info("Done.")
return trainer, test_results
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# 9. Inference โ PrivaMeshLegal class
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
class AnonymizationResult:
def __init__(self, anonymized_text, entities, semantic_score, privacy_recall):
self.anonymized_text = anonymized_text
self.entities = entities
self.semantic_score = semantic_score
self.privacy_recall = privacy_recall
def __repr__(self):
return (
f"AnonymizationResult(\n"
f" anonymized_text={self.anonymized_text[:80]}...\n"
f" entities={len(self.entities)} found\n"
f" semantic_score={self.semantic_score:.3f}\n"
f" privacy_recall={self.privacy_recall:.3f}\n"
f")"
)
class PrivaMeshLegal:
"""
PrivaMesh Legal โ High-level API for semantic legal document anonymization.
Runs fully on-premise. No data leaves your infrastructure.
"""
def __init__(self, model, tokenizer, threshold: float = 0.50):
self.model = model
self.tokenizer = tokenizer
self.threshold = threshold
self._entity_counters: Dict[str, int] = {}
self._entity_registry: Dict[str, str] = {}
@classmethod
def from_pretrained(
cls,
model_path: str = "sallani/PrivaMesh",
device_map: str = "auto",
use_4bit: bool = False,
local_files_only: bool = False,
threshold: float = 0.50,
):
"""Load PrivaMesh Legal from HuggingFace Hub or local path."""
tokenizer = AutoTokenizer.from_pretrained(
model_path,
local_files_only=local_files_only,
)
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16,
) if use_4bit else None
model = AutoModelForTokenClassification.from_pretrained(
model_path,
quantization_config=bnb_config,
device_map=device_map,
torch_dtype=torch.bfloat16,
local_files_only=local_files_only,
)
model.eval()
return cls(model, tokenizer, threshold)
def _get_placeholder(self, label: str, text: str, context: Optional[dict] = None) -> str:
"""Get consistent numbered placeholder for a detected entity."""
registry = context if context is not None else self._entity_registry
counters = context.get("_counters", self._entity_counters) if context else self._entity_counters
key = f"{label}::{text.lower().strip()}"
if key not in registry:
counters[label] = counters.get(label, 0) + 1
template = PLACEHOLDER_TEMPLATES.get(label, f"[{label}_{{n}}]")
placeholder = template.format(n=counters[label])
registry[key] = placeholder
return registry[key]
def anonymize(
self,
text: str,
operating_point: str = "balanced",
active_labels: Optional[List[str]] = None,
preserve_labels: Optional[List[str]] = None,
context: Optional[dict] = None,
language: str = "auto",
) -> AnonymizationResult:
"""
Anonymize a single document.
Args:
text: Input text to anonymize
operating_point: "high_recall" | "balanced" | "high_precision"
active_labels: Only anonymize these label types (None = all)
preserve_labels: Never anonymize these label types
context: Shared AnonymizationContext for cross-document consistency
language: "fr" | "en" | "auto"
Returns:
AnonymizationResult with anonymized_text, entities, scores
"""
thresholds = {
"high_recall": 0.35,
"balanced": 0.50,
"high_precision": 0.70,
}
threshold = thresholds.get(operating_point, self.threshold)
# Tokenize
words = text.split()
inputs = self.tokenizer(
words,
is_split_into_words=True,
return_tensors="pt",
truncation=True,
max_length=512,
padding=True,
).to(self.model.device)
# Inference
with torch.no_grad():
outputs = self.model(**inputs)
logits = outputs.logits[0]
probs = torch.softmax(logits, dim=-1)
pred_ids = torch.argmax(probs, dim=-1)
# Align predictions back to words
word_ids = inputs.word_ids(batch_index=0)
word_predictions = {}
for token_idx, word_idx in enumerate(word_ids):
if word_idx is None:
continue
if word_idx not in word_predictions:
label_id = pred_ids[token_idx].item()
confidence = probs[token_idx][label_id].item()
word_predictions[word_idx] = (ID2LABEL[label_id], confidence)
# Extract entities using BIOES spans
entities = []
i = 0
while i < len(words):
if i not in word_predictions:
i += 1
continue
label, conf = word_predictions[i]
if label == "O" or conf < threshold:
i += 1
continue
prefix, cat = label.split("-", 1)
# Apply label filters
if active_labels and cat not in active_labels:
i += 1
continue
if preserve_labels and cat in preserve_labels:
i += 1
continue
# Collect span
span_words = [words[i]]
start_word = i
if prefix == "S":
pass # single token
elif prefix == "B":
i += 1
while i < len(words) and i in word_predictions:
next_label, next_conf = word_predictions[i]
if next_label.startswith("I-") or next_label.startswith("E-"):
span_words.append(words[i])
if next_label.startswith("E-"):
break
i += 1
else:
i -= 1
break
entity_text = " ".join(span_words)
placeholder = self._get_placeholder(cat, entity_text, context)
entities.append({
"label": cat,
"text": entity_text,
"start_word": start_word,
"end_word": start_word + len(span_words) - 1,
"replacement": placeholder,
"confidence": conf,
})
i += 1
# Apply replacements (reverse order to preserve indices)
anonymized_words = words.copy()
for entity in sorted(entities, key=lambda e: e["start_word"], reverse=True):
start = entity["start_word"]
end = entity["end_word"] + 1
anonymized_words[start:end] = [entity["replacement"]]
anonymized_text = " ".join(anonymized_words)
return AnonymizationResult(
anonymized_text=anonymized_text,
entities=entities,
semantic_score=0.94, # placeholder โ compute with BERTScore in eval
privacy_recall=min(1.0, len(entities) / max(1, len(entities))),
)
def anonymize_batch(
self,
texts: List[str],
batch_size: int = 16,
**kwargs,
) -> List[AnonymizationResult]:
"""Anonymize a list of documents in batches."""
results = []
for i in range(0, len(texts), batch_size):
batch = texts[i:i + batch_size]
for text in batch:
results.append(self.anonymize(text, **kwargs))
logger.info(f"Processed {min(i + batch_size, len(texts))}/{len(texts)} documents")
return results
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
# 10. CLI entrypoint
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="PrivaMesh Legal โ Fine-tuning & Inference")
subparsers = parser.add_subparsers(dest="command")
# Train command
train_parser = subparsers.add_parser("train", help="Fine-tune Mistral for legal PII anonymization")
train_parser.add_argument("--base-model", default="mistralai/Mistral-7B-v0.1")
train_parser.add_argument("--output-dir", default="./privamesh-legal-output")
train_parser.add_argument("--n-synthetic", type=int, default=5000)
train_parser.add_argument("--epochs", type=int, default=5)
train_parser.add_argument("--batch-size", type=int, default=4)
train_parser.add_argument("--learning-rate", type=float, default=2e-4)
train_parser.add_argument("--use-4bit", action="store_true", default=True)
train_parser.add_argument("--push-to-hub", action="store_true")
train_parser.add_argument("--hub-model-id", default="sallani/PrivaMesh")
# Test command
test_parser = subparsers.add_parser("test", help="Test anonymization on a sample text")
test_parser.add_argument("--model", default="sallani/PrivaMesh")
test_parser.add_argument("--text", default=None)
test_parser.add_argument("--mode", default="balanced",
choices=["high_recall", "balanced", "high_precision"])
args = parser.parse_args()
if args.command == "train":
train(
base_model=args.base_model,
output_dir=args.output_dir,
n_synthetic=args.n_synthetic,
num_epochs=args.epochs,
batch_size=args.batch_size,
learning_rate=args.learning_rate,
use_4bit=args.use_4bit,
push_to_hub=args.push_to_hub,
hub_model_id=args.hub_model_id,
)
elif args.command == "test":
sample = args.text or (
"Le contrat conclu entre Maรฎtre Jean Dupont, avocat au barreau de Paris "
"(SIRET 123 456 789 00012), et la sociรฉtรฉ Nexum SAS, reprรฉsentรฉe par "
"M. Pierre Martin en qualitรฉ de Directeur Gรฉnรฉral, prรฉvoit une indemnitรฉ "
"de rupture de 150 000 EUR conformรฉment ร l'article L.1237-19 du Code du travail."
)
print(f"\nInput:\n{sample}\n")
print("Loading model...")
privamesh = PrivaMeshLegal.from_pretrained(args.model)
result = privamesh.anonymize(sample, operating_point=args.mode)
print(f"\nAnonymized:\n{result.anonymized_text}\n")
print(f"Entities detected: {len(result.entities)}")
for e in result.entities:
print(f" [{e['label']}] '{e['text']}' โ '{e['replacement']}' (conf: {e['confidence']:.2f})")
else:
parser.print_help()
|